Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more

Jensen Huang argued that the AI buildout is supported by a much broader compute economy than the revenue of a handful of AI labs. The interviewer said Huang’s forecast of at least $2 trillion in cumulative compute investment over five years had already been exceeded, and estimated AI labs’ annual revenue at about $200 billion. Huang cautioned against multiplying revenue per gigawatt by planned capacity or treating lab revenue as the sole test of payback: infrastructure also serves general-purpose computing, cloud providers, neoclouds, sovereigns, enterprises, startups and many labs. He agreed with a proposed breakdown of next year’s 40-gigawatt buildout into roughly thirds for general-purpose replacement, neoclouds and other users, and labs.

Huang expects global infrastructure spending to approach $3 trillion to $4 trillion annually by decade’s end, closer to $4 trillion, including about $1.5 trillion for compute. He said compute infrastructure spending had been about $500 billion and argued that, with Moore’s Law over, doubling computing capacity now costs roughly twice as much. As a comparison, he put general-purpose compute at $5 billion per gigawatt and Vera Rubin at $50 billion, asking whether the latter would deliver ten times the value. He said annual infrastructure building and refreshing could reach $4 trillion to $5 trillion, and that a $10 trillion annual run rate would not surprise him.

Huang said inference compute could grow by a billion times because reasoning requires generating options, planning, assessing outcomes and reflecting, rather than simply repeating memorized information. He described models grounding answers with search and harnesses adding working memory, extracted knowledge and knowledge graphs. Teaching agents to use existing tools is another major unlock: coding is an early on-ramp, while design, creative work and science could follow. He described the singularity as agents collaborating on long-running tasks and keeping one another on task. More capable systems assigned more work will compound compute demand, even as efficiency and added supply lower the cost of intelligence; Huang’s goal is for that cost to fall by a factor of two or 20 per year.

The Navier–Stokes effort illustrated the cost and potential of agentic research: the interviewer cited 10,000 agents and an estimated $4 million in compute, with comparable work potentially costing five to ten times less within a few years. Huang said AI can build on established scientific and mathematical work to explore large search spaces. He also sees physical AI as a growing business: NVIDIA generated $10 billion in the first six months of the year, following $10 billion the prior year and $5 billion the year before. The field combines perception, reasoning, planning and action across robotics, vehicles, manufacturing, logistics and healthcare. Huang said it needs training computers, digital twins such as Omniverse, robots and complete software stacks, and called the proposed $100 billion robotics market potentially too small.

Huang identified land, power, permitting, skilled construction labor and data-center shells as near-term constraints, with land, power and buildout resources likely among the most important bottlenecks next year. The interviewer cited estimates of 50 gigawatts of announced chips versus 35–40 gigawatts of compute that available infrastructure might support. Huang said NVIDIA remains supply-constrained despite a large order book and demand well above guidance for 70% revenue growth; substrates and package materials, including organic materials affected by heat and stacking, are also difficult to scale, as is silicon photonics. He said the supply gap could remain hard to close for a few years. NVIDIA’s tokens-per-watt performance matters when power is scarce, and Huang said its share is increasing.

Huang said about 80% of startups use open models, often combining them with proprietary data and customized systems for specific sectors; he expects general-purpose closed labs to pursue AGI while domain-focused firms seek superintelligence in particular fields. He said NVIDIA’s “open shelf” announcement drew more than 100 companies. On safety, he argued that capability, safety and security should advance together: autonomous AI should start with zero trust, minimum access and tight containment, with permissions expanded as confidence grows, plus out-of-band monitoring to detect violations and escalate or shut systems down. He backed rapid, actionable safeguards over waiting for a perfect solution, and said safety and security are existential for the industry because AI must earn the chance to demonstrate its value.